Video Conform
Portability: one stdlib script driving ffmpeg. Without --run it only prints the plan and command.
Conform is where a good cut gets rejected: wrong loudness, music drowning the
voice, no captions, a file that will not play on a phone. The script makes one
ffmpeg pass and then checks the result instead of trusting it.
Workflow
C=${CLAUDE_PLUGIN_ROOT}/skills/video-conform/scripts/conform.py
python3 $C --render film/frames/render.json --scene film/scene.json --vo film/vo.wav --music film/bed.mp3 --out film/final.mp4 # plan
python3 $C --render film/frames/render.json --scene film/scene.json --vo film/vo.wav --music film/bed.mp3 --out film/final.mp4 --run
python3 $C --video film/cut.mp4 --scene film/scene.json --loudness music --burn-captions --out film/final.mp4 --run
Music is mixed about 12 dB under the voice. captions.srt is always written
when a scene is given (two lines of up to 42 characters per cue; a long voice-over line is split
into several cues across its shot, never cut);
--burn-captions also draws them in, which most social feeds need because
they autoplay muted.
QC after encode: duration within one frame of the scene and loudness within
1 LU of the target. Loudness is single-pass loudnorm, which lands within
1 LU on clips of about 5 s and longer; on very short stings measure and re-run
rather than widening the gate. A voice-over longer than the picture is warned
about, and a render with missing frames is refused. A miss is exit 4; fix the input (a clipped VO, a silent
track) rather than re-running until it passes.
Delivery checklist
See delivery_specs.md for sources.